jarvis-consolidate

Memory Consolidation

LLM-driven consolidation finds groups of redundant memories and synthesizes them into authoritative summaries.

Commands

1. Dry Run (default)

Show what would be consolidated without making changes.

/jarvis-consolidate
/jarvis-consolidate dry-run

Output: List of memory clusters with size, similarity, importance, and previews.

2. Interactive

Step through each cluster, approve or reject consolidation.

/jarvis-consolidate interactive

For each cluster:

1. Show memory contents and similarity score
2. Call LLM to generate consolidated summary
3. Show summary + confidence score + any contradictions
4. Ask user: approve / reject / edit

3. Undo

Reverse a consolidation run.

/jarvis-consolidate undo <run_id>

Restores all superseded originals and soft-deletes the consolidated summaries.

Workflow

1. Call find_consolidation_candidates() — ANN-based clustering
2. For each cluster: a. Load full contents b. Run secret scanner on cluster contents c. Build consolidation prompt, call LLM d. Parse response, assess confidence e. Route based on mode (dry-run / interactive / auto)
3. For approved consolidations: transactional apply
4. Report results

Safety

• Contradictions flagged, never auto-resolved — human reviews via interactive mode
• Originals preserved — superseded, not deleted
• Reversible — undo by consolidation_run_id
• Secret scanning — cluster contents checked before LLM call
• Disabled by default — memory.consolidation.enabled: false